Topic: 50-core Intel Xeon Phi of no use to us? (Read 14306 times)

That was on an OpenCL miner.  I figured it would be closer to the max potential than that.  Right now I use my cpu for litecoin anyway, but at those numbers I may switch; probably not worth it even at 66 MH/s, though.

Also, sorry for the bad math.  I've edited the original post where I put it.  Thanks for the insight.

The problem with the equation that Koooooj posted is that:

1) Those ops are for an OpenCL kernel written for a GPU (most likely HD5xxx VLIW), not the x86-derived RISC operations in the 3930K.
2) Modern day processors are superscalar, retiring much more than 1 instruction per clock cycle.

Maybe thats on one core? https://en.bitcoin.it/wiki/Mining_hardware_comparison#Intel

Shows the 3930k as getting 66MH/s at 4.6GHz at 180Watts.

What miner are you using? With the Ufasoft miner I remember seeing something like of 25 mhash/s on a 6-core Xeon clocked less than your i7.

I was afraid my math was flawed.  I will maintain that if both the Phi and my 3930K both use SIMD and MIMD, or at least vaguely similar architecture (I have read that the Phi will use a pentium like architecture, just built on 22nm and tied together for better parallel processing), then the (MHash/s) per (core count * clock speed) comparison should at least be ballpark, and the validation that the MHash/s that I see running with a CPU miner is at least ball park lends at least a little bit of weight.  Not so much weight that it should be taken as truth without thought, but enough that it should perhaps be considered.

Ignoring that implication, though, there is the simple fact that this is going to be a VERY expensive card.  With well-proven GPUs easily available that will probably compete on the MHash/s, FPGAs readily buyable from a few manufacturers, and the ASICs in development, buying a card that will probably be priced a little above the nVIDIA Tesla cards (since they have x86 support) will simply not be economical for bitcoin miners.  For people building massive super-computing clusters for general computation, it is a different story.

I think the technology you are looking for here is superscalar. Modern day CPUs have IPC well over 1 which is only possible due to the architecture being superscalar. Superscalar means that the processor can issue and retire more than 1 instruction per cycle.

Hyperthreading (aka simultaneous multithreading) is simply a technology that simply tries to keep the processor as busy as possible by simultaneously assigning instructions from different threads to fill up the functional units as efficiently as possible.

should point out that your math is crap, since your core i7 has SIMD and MIMD extensions that do far more than one operation per second, as the name suggests(Single input, multiple data). Besides that, Hyperthreading doesnt make your cpu run more instructions per second, it just allows for better parallelization. You can also bet that the Phi will have SIMD and MIMD instructions. Then it should be added that a cpu will not reach its nominal ops/second speed in most scenarios(bitcoin mining might be close to the theoretical values because the functions are rather compact and can be held within the 1:1 L1 cache of the cpu most of the time) du to the limited L1 and L2 cache, and other tasks/kernel messing with the instruction pipeline. finally, not every instruction can be done in one clock. then again some instructions can be done in less than one effective clock(strict prefetching).

Just saw the reference to https://bitcointalksearch.org/topic/m.974306.

Probably more accurate than my math.

(EDIT) PAY NO ATTENTION TO ME. I AM APPARENTLY A COMPLETE IDIOT.  THANKS FOR SETTING ME STRAIGHT.

Rather than suggesting numbers that I think sound good, I'll attempt to recreate a lot of the math found in https://bitcointalksearch.org/topic/m.996572

In that thread, it was stated that the Bitcoin Hash Function takes about 3375 operations.  If we can assume that each of the 50 cores can perform one operation per clock, and we assume an aggressive clock of, say, 1,500 MHz (1.5 GHz), then the following math ensues:


Now, I'm not saying that that is what the actual performance will be, but it is a number that is supported by real math.  If you want, you can play around with those numbers to get different results, but I doubt that it'll wind up getting all the way up to 2,400 MHash/s.  Note that you aren't allowed to change the 50 (there will definitely be 50 cores), and you aren't allowed to change the 3375--that's just the amount of math it takes to do a hash.  All you can do is speculate on the clock speed and the number of operations that can be done by a processor in a given clock cycle.

To give a comparison to show the validity of the math, I have an Intel Core i7 3930K @ 4.4GHz.  That's:

Compare this to the 8.7 that I actually see.  The discrepancy is caused by the hyperthreading that allows more than one operation to be done in a given clock cycle, since there are two concurrent threads executing.  Still, though, even with a hyperthreaded processor the math is pretty close.

In all, it's not too surprising that it's not that high, as it's x86, not straight OpenCL.  It's good for more things, but not as good at the few things that an ATI/AMD graphics card (or FPGA/ASIC) is good at.

No

We need exactly 0 float point calculations. And the Xeon Phi will have a lot of float calculation power, it will have 1TFLOPs in Double Precision, all with x86 code (easier to use than opencl or cuda)

Never heard of a GPU too? Because a GPU is a co-processor 

Stop the mystery. It is me

What is the price for this monster anyway ? No doubt like $2000 or something because this is not for users like us but big business.

thats right you don't need approval, and because we guys have no clue, we don't care

i dont't know what knowledge about the community has to do with knowloedge about gpu/cpu architecture ?

please explain....

Everyone is entitled to their own opinion.  I know so much about the bitcoin community ... you guys have no clue.  I have no reason lie.  I don't need the approval of anyone from this forum.

That I do not know. 

2.4GHash/s but what hashes? bitcoin double sha2-256 or maybe 2.4 md5 GHash/s? This is important.

I think you "friend" who works for Intel is off by a factor of 10. 240 Mhash/s might be possible, but not 2400 Mhash/s.
Some have predicted a bit more than 240Mhash/s of course, but not off by much. Numbers just don't add up to do as you suggest.
Mrb post details the core of those numbers of how 2400 Mhash/s isn't going to be possible.

It's all mute now with FPGA coming out at lower prices and better results without using tons of electricity.
ASIC is of course on the horizon, should prove to be an improvement that most FPGA out right now.
I don't expect what BFL suggests, but ASIC will eventually be the next step forward, after FPGA.

I don't think you are telling the truth.


Xeon Phi should get about 280 Mhash/s (1200 billion 32-bit operations per sec / 4300 operations):
https://bitcointalksearch.org/topic/m.974306

By the time this becomes available, it better be 2.4 GH/s for $100 to compete with a 3.5GH/s ASIC for $150.

ok i want to shit my pants! who is it ? 

I actually know a person who works for Intel who is developing this.  You would all shit your pants if you knew who he was in this community and you probably wouldn't believe me.

He says this is going to get about 2.4 Ghash range (they are still testing it) and it was supposed to cost about 1000 dollars but maybe they want to compete with asic and offer cheaper prices?  But again, this card isn't really for Bitcoin, so who knows.